Does the RTX PRO 6000 have NVLink? How its GPUs talk to each other

The RTX PRO 6000 Blackwell has no NVLink: its GPUs talk over PCIe 5.0 x16. What that means for multi-GPU inference, tensor parallelism and training.

Row of IBM cabinets of the Summit supercomputer at Oak Ridge National Laboratory, whose NVIDIA Volta GPUs were linked with NVLink
Photo: ORNL History (Summit supercomputer) by Oak Ridge National Laboratory, Wikimedia Commons, CC BY 2.0

No. The NVIDIA RTX PRO 6000 Blackwell has no NVLink, in any edition. Lenovo's product guide for the Server Edition lists "NVLink support: No" and a host interface of PCIe 5.0 x16. When several RTX PRO 6000 cards work together, every byte between them goes over PCIe. That is fine for some jobs and a real limit for others. Below is where the line falls, with published numbers.

Which setup for whom

  • One model that fits on one card (up to about 96 GB with its cache) → one GPU, no interconnect needed. A 70B model in 4-bit, a 30B model in FP8 or gpt-oss-120b fit on a single card. NVLink would change nothing here.
  • Many users, a model that fits on one or two cards → several copies side by side. Each copy runs on its own card or pair, and a load balancer spreads the requests. PCIe is barely used, so this scales almost linearly.
  • One very large model split across four to eight GPUs, lowest latency → NVLink hardware (H100, H200, B200). Tensor parallelism sends data between GPUs at every layer, and PCIe becomes the bottleneck.
  • Training or fine-tuning across many GPUs → depends on size. LoRA on one or two cards works well. Full training of large models across eight GPUs favours NVLink systems.
  • Serving several models or many users on 768 GB of GPU memory → a whole 8-GPU server. Our server in Sofia has eight RTX PRO 6000 Blackwell Server Edition cards. It rents from $1.80 per GPU-hour, or about $10,510 a month for the whole machine. It is rented into early 2027 and we take reservations for the period after. See the server.

What the specifications say

RTX PRO 6000 Blackwell Server Edition H100 / H200 (SXM) B200
GPU-to-GPU link PCIe 5.0 x16 NVLink, 4th generation NVLink, 5th generation
Bandwidth per GPU about 128 GB/s (both directions together) 900 GB/s 1,800 GB/s
Memory 96 GB GDDR7, 1,597 GB/s 80 / 141 GB HBM3 / HBM3e HBM3e

NVIDIA's own page for the Server Edition lists 96 GB of GDDR7 at 1,597 GB/s, up to 600 W and PCI Express Gen 5. It does not list NVLink at all. The NVLink figures per GPU come from NVIDIA's NVLink page: 900 GB/s for the fourth generation (Hopper) and 1,800 GB/s for the fifth (Blackwell data centre GPUs). PCIe 5.0 x16 carries about 64 GB/s in each direction.

So an H100 SXM can exchange data with its neighbours about seven times faster than an RTX PRO 6000, and a B200 about fourteen times faster.

The workstation cards used to have a bridge

Older professional cards did have NVLink. NVIDIA's datasheet for the RTX A6000 (Ampere, 2020) lists an NVLink bridge for two cards with up to 112 GB/s of bidirectional bandwidth. The RTX PRO 6000 Blackwell family has no bridge connector, so this option is gone. On today's cards, PCIe 5.0 x16 alone (about 128 GB/s in total) is roughly the same bandwidth that the old A6000 bridge gave, but it is shared with all other traffic to the CPU, storage and network.

What it means in practice

The amount of data that crosses the link depends on how you split the work.

Tensor parallelism splits every layer of the model across GPUs. After each layer the GPUs combine partial results, so they talk constantly. This is the mode where the link matters most.

Pipeline parallelism gives each GPU a block of layers. Data crosses the link only between blocks, much less often, at the cost of some latency.

Data parallelism or replicas runs a full copy of the model on each GPU or group. For inference the GPUs do not talk to each other at all.

A public benchmark shows the effect. In November 2025 CloudRift tested the RTX PRO 6000 against H100 and H200 with vLLM:

Setup (CloudRift, vLLM) RTX PRO 6000 H100 SXM H200
1 GPU, GLM-4.5-Air, 4-bit 3,140 tokens/s 2,987 tokens/s not tested
4 GPUs, tensor parallel, Qwen3-Coder-480B, 4-bit about 1,600 tokens/s about 2,100 tokens/s about 3,400 tokens/s
8 GPUs, tensor parallel, GLM-4.6, FP8 about 900 tokens/s about 2,700 tokens/s about 3,600 tokens/s

On one GPU the RTX PRO 6000 was slightly ahead of the H100. With one model split across eight GPUs, the H100 was three times faster. In the same test, eight independent copies of the smaller model on eight RTX PRO 6000 cards reached 15,713 tokens per second. Our reading: the gap at eight GPUs comes largely from the link, while the replica result shows how well the cards scale when they do not need to talk.

How to get the most out of PCIe

  • Prefer replicas over splitting. If a model fits on one or two cards, run several copies instead of one copy across eight. vLLM and SGLang both support this. See vLLM vs Ollama vs SGLang.
  • Keep tensor parallel groups small. A tensor parallel size of 2 or 4 usually costs far less than 8. For larger models, combine tensor parallelism inside a group with pipeline parallelism between groups.
  • Use FP8 or 4-bit weights. Smaller weights mean a model fits on fewer cards, so less traffic crosses PCIe. Our VRAM calculation shows how many cards a model needs.
  • Check the link. nvidia-smi topo -m shows how the GPUs connect, and nvidia-smi -q shows the current PCIe generation and width. A card running at Gen4 or x8 by mistake halves the bandwidth.
  • Test peer-to-peer before you rely on it. Direct GPU-to-GPU transfers over PCIe depend on the platform, BIOS settings (ACS, IOMMU) and virtualisation. Users on NVIDIA's developer forum report problems with P2P in some passthrough setups. NCCL's own tests show quickly whether it works on your machine.

Choose an NVLink system (HGX H100, H200 or B200) if you train large models from scratch across eight GPUs, or serve one model too big for four cards to many users at low latency. These systems cost much more per GPU-hour, so it pays to check first whether replicas or a smaller tensor parallel group would do. For a side-by-side of the cards, see B300 vs H200 vs RTX PRO 6000.

Questions

Can I add an NVLink bridge to the RTX PRO 6000? No. The cards have no NVLink connector, and Lenovo's product guide lists NVLink support as "No".

Does the RTX PRO 6000 pool its memory across cards? Not in hardware. Frameworks such as vLLM split the model so that each GPU holds part of it, which lets eight cards serve a model of up to about 768 GB in total, minus the cache. Every exchange between the parts goes over PCIe.

Is MIG a replacement for NVLink? No, it is the opposite. MIG splits one GPU into up to four isolated parts. It helps when your models are small, not when they are big. See NVIDIA MIG on the RTX PRO 6000.

Is PCIe 5.0 enough for fine-tuning? For LoRA and QLoRA on one or two cards, usually yes. For full fine-tuning of large models across eight GPUs, the gradient exchange makes NVLink systems faster.

Sources

SHARE